The Quarter Before the Numbers Say So
Schleswig-Holstein's company insolvencies rose 36.6% in one quarter. Building permits collapsed two years earlier. And the signal everyone treats as leading turned out to lag.
Between April and June 2026, 291 companies in Schleswig-Holstein filed for insolvency. A year earlier, over the same three months, it was 213.
That is +36.6%, and it is the third alert in that state since the turn of the year: +28.6% for November–January, +31.8% for December–February, +36.6% for April–June. The trend is not a spike. It is escalating.
All of it comes from statistics anyone can download for free. The question this namespace exists to answer is what you have to build around that data before it can tell you something in time to act on.
The short version: monthly, state-level early warning from public German statistics — 5 sources, 15 detectors, 234 findings, 6 investigations. Company insolvencies, unemployment, business registrations and deregistrations, building permits, births and deaths, plus the statistical office’s own change feed and the EU legal framework underneath it all. One of its most useful results is a hypothesis it refuted.
What an alert has to survive to be worth sending
The naive version of this is easy and useless: compare the latest month to the same month a year ago, alert when it moves. You get an alert every month, in every state, forever.
Three design decisions make the difference.
Three-month windows, not months. Monthly insolvency counts in a small state are noisy — Bremen runs in the dozens. Rolling three-month windows against the same window a year earlier smooth that without hiding a turn.
A floor under the base. A rule with no minimum fires on 4 → 7 and calls it +75%. Alerts require a base of at least 60 in the comparison window, so a percentage means something.
Alerts are onsets, not states. This is the one that matters most. The first version tagged every month a rule held — and produced 252 “natural balance” findings, because births have fallen every year since 2021 and every window was a new low. Now a finding is kept only on the month a streak begins. An alert that started in June stays on June when July arrives, instead of resolving because the window moved on. An eight-month rise is one finding, not eight, and the running streak sits in the month’s metadata.
There is a subtlety in that rule worth stating, because it is where an early-warning system usually goes wrong: an escalation inside an existing streak has to be its own finding. Schleswig-Holstein’s +36.6% would have been silent — swallowed by a streak that began in January at +28.6% — if the severe band were not tracked separately from the ordinary one.
The signal everyone gets backwards
Here is the finding we would send to a credit or procurement team first, and it is a negative result.
In January 2024, unemployment rose by at least half a percentage point year on year in five German states in the same month. Insolvency jumps in those same states arrived two to three months later.
The namespace wrote that down as a testable claim in the opposite direction — “unemployment turns follow insolvency jumps within six months” — and recorded it as refuted.
That matters because insolvency statistics are widely treated as the leading indicator of regional distress. In this data they are not. They are the confirmation. The labour market moved first.
A second hypothesis in the same case — that the end-2024 wave of business deregistrations leads 2025 insolvencies — sits at proposed, with the test written and waiting for data.
And the shape of the January 2024 event is itself a finding: five states simultaneously is a national step, not five regional shocks, which changes what you do about it.
Two years of warning in the construction sector
The longest lead time in the namespace comes from building permits.
Permits for dwellings collapsed across most German states at once in early 2024. Construction then appears in the sector-level insolvency alerts one to two years afterwards.
The mechanism is not mysterious — a permit is a commitment to build, and its absence works through the order books of everyone downstream — but the lag is long enough to be genuinely actionable for anyone with construction exposure.
Getting the permits data at all took work worth mentioning. Table 31111-0120 is
too large to request in one call and has to be fetched per state. Its totals sit
under codes that the documented “give me the total” query does not return —
BTK-GEB and GBD-W-NW — so a query written exactly as documented comes back
empty. That trap is recorded as its own finding.
A third hypothesis notes that the slump is not over everywhere: supported, which is the part a 2024-vintage analysis would miss.
Watching the ground the alerts stand on
Two of the six investigations are not about the economy at all. They watch the data supply itself, which is what separates a system from a script.
Publication lag. The statistical office’s change feed says when each watched statistic was updated. Lags differ by an order of magnitude: unemployment arrives three days before its month ends, deaths at 15 days, business registrations at 42, migration at 70, insolvencies at 73, births at 76.
If you build a monthly risk review, that table is the schedule. There is no point asking about insolvencies in week one.
A standing question watches for anything arriving more than twice as late as its own median — a plausible precursor to a revision or a withdrawal. It is empty today, and that is the correct state: the median is built from lag history the feed accumulates across runs, and nothing has been unusually late. The question says so on its face, so an operator reading “0 matches” does not mistake silence for breakage.
The legal framework. The EU rules behind insolvency and business statistics changed repeatedly across 2025–2026, and the Court of Justice keeps refining cross-border insolvency. Both are tracked as findings on the acts themselves, because a definition change is indistinguishable from a trend change if you only watch the numbers.
One rule, two severities
A small change with an outsized effect on how usable the alerts are.
The insolvency rule has two bands: medium at +20%, high at +30%. That used to require two separate detectors with duplicated descriptions, and every standing question and every case had to name both.
Now one detector carries the rule and each finding carries its own band. The question “insolvency jumps” names one thing. A reader sees severity on the finding, where it belongs, rather than inferring it from which of two near-identical detectors fired.
The constraint that makes this safe: a connector may lower a finding’s severity within what the detector allows, and may never raise it past it. The detector is where the operator’s decision lives.
What Classifyre contributed
Alerts arrive without anyone opening a page. Each standing question raises a notification when a run produces new matches — including on questions nobody has opened yet. An insolvency onset, a permit collapse, a withdrawn table each surface on their own.
Every alert carries its arithmetic. A finding is not “Schleswig-Holstein:
risk up”. It is 2026-04..2026-06: +36.6% — 291 vs 213 Unternehmensinsolvenzen,
with a lineage edge back to the exact table and measure code, and the sector
split beside it.
The investigation holds its own disconfirmation. The Schleswig-Holstein case has five hypotheses. “One or two sectors carry the rise” is refuted — the rise is broad. “Construction led the February escalation” is supported. “It is not a single large group filing” is proposed, with the test written down. That is a document a credit committee can argue with.
The sources run themselves. Monthly indicators on a weekly cron, the change feed daily, the legal sources weekly. Windows missed while the instance is down are caught up rather than silently skipped.
What it will not tell you
Stated plainly, because an early-warning system that oversells itself is worse than none:
- It is state-level, not company-level. It tells you Schleswig-Holstein’s construction sector is deteriorating. It does not name your counterparty.
- Insolvency statistics lag. The namespace’s own refuted hypothesis says so.
- Three months is the resolution. A shock inside one month is not visible.
- The unusually-late signal needs history. It cannot fire until the feed has been watched long enough to have a median to compare against.
Used as one input among several — with the publication-lag table telling you when each number is actually worth asking about — it does something a quarterly report cannot: it says which state, which sector, which month it started, and whether it is still escalating.
Who this is for
Credit and counterparty risk, trade credit insurance, supply-chain and supplier monitoring, regional banking, construction and building materials, staffing, and public-sector economic development.
Two things transfer to any early-warning problem regardless of domain: alert on the start of a streak, not on every month it holds, and write down the lead-lag relationship you believe in and test it, because the one in this data ran the opposite way to the consensus.
See it yourself
The namespace is live, with every source, detector, standing question, case and lineage edge described above:
showcase.classifyre.com/de-fruehwarnung
Start with Arbeitsmarkt: Januar 2024 kippte in fünf Ländern zugleich and read the refuted hypothesis — then follow Schleswig-Holstein’s three alerts back to the monthly table they were computed from.
More case files
Four Hundred Districts, Ten Years: Telling Real Growth From the Kind That Isn't
Three districts around Berlin grow far faster than their structure predicts. Nine look like they are under housing pressure — seven of them are not. A regional picture of Germany where every number comes with what it can and cannot support.
Köln Is Not Shrinking: When the Official Number Is Not the Number
Germany's official population figures say 44 districts shrank over the last decade. Once the 2022 census correction is removed, those same districts grew. Here is the namespace that tells the two apart — and what else it found hiding in plain sight.
Who Am I Actually Dealing With? Austria's Company Register, Made Answerable
Two Austrian companies, both registered in 1992, both "active". One filed 21 sets of accounts, the other has filed none in 34 years. Here is how we built the namespace that tells them apart.